All releases
v0.4.0

Pull request review, auto-fix, and dictation

Review GitHub pull requests, bring failed checks into the agent composer, enable optional PR auto-fix, and dictate a prompt. This release also improves first-run setup, thread status, and Windows reliability.

A workspace for GitHub pull requests

Open the Pull Requests page to review summaries, code, comments, and check results across projects. Send a review finding or failed check to the linked thread's composer with its PR context attached. A row above the composer keeps PR and CI status visible while you work.

Optional auto-fix for new PR feedback

Enable auto-fix on a thread to ask its agent to handle new failed checks and review comments on an open GitHub PR. The watcher checks about every two minutes, skips busy or waiting threads, and allows up to three automatic turns per thread and PR during a server session. It starts from new feedback after activation and does not guarantee a fix or merge.

English dictation in the composer

Use the microphone to turn speech into a draft after downloading a transcription model. Processing runs on the connected environment's server. With a local environment that is your machine; with a remote environment, audio goes to that server.

A clearer first run on Windows

See which coding agents and Git tools are ready, install missing tools, and sign in from setup. Newly installed tools are detected without restarting the app. One signed-in agent is enough to start your first thread, and GitHub repository browsing includes your organizations.

PR merge state and thread wrap-up stay together

See GitHub merge queue and auto-merge state, with actions that follow the repository's rules. Choose whether a thread wraps up after its PR merges, using a per-thread setting or the app-wide default.

Also improved

  • Long database migrations no longer look like a failed desktop startup. History backfills avoid repeated scans, and interrupted sessions recover more predictably.
  • Windows worktrees handle long paths and cleanup failures more reliably. New worktrees can use a refreshed upstream base when the local branch is behind.
  • Finished subagents stop showing as Running, their original prompts remain visible, and non-blocking Codex questions stay answerable while work continues.
  • Images referenced or returned by agents render inline, including Claude image output. Missing files keep their file link.
  • Press Tab in an empty composer to accept a Claude prompt suggestion. Long suggestions are easier to read.
  • The latest task plan stays available in long conversations after reloading, with clearer progress and step counts in the activity panel.